April 2025 monthly summary for liguodongiot/transformers focusing on delivering a critical bug fix that enhances inference reliability under tensor parallelism. Key work centered on correcting the Qwen model's attention reshaping logic to ensure proper output shapes during inference, eliminating a class of runtime errors.

February 2025 monthly summary for InternLM/InternEvo focused on MoE stability and training correctness. Delivered a targeted bug fix to MoE activation, addressing late-release behavior by refining handling of empty inputs and ensuring correct gradient calculations within Mixture-of-Experts layers. Additionally, cleaned up management of auxiliary loss values inside MoE components to improve stability in edge cases. These improvements reduce training disruptions and enhance reliability for large-scale MoE deployments, supporting more robust model updates and experimentation.